Fence propagation from SYCL

Separate rankSum from sum

Grid/DisableWarnings.h

@@ -45,7 +45,7 @@ directory
  //disables nvcc specific warning in json.hpp
 #pragma clang diagnostic ignored "-Wdeprecated-register"
 
-#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
+#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
  //disables nvcc specific warning in json.hpp
 #pragma nv_diag_suppress unsigned_compare_with_zero
 #pragma nv_diag_suppress cast_to_qualified_type

Grid/Grid_Eigen_Dense.h

@@ -14,7 +14,7 @@
 /* NVCC save and restore compile environment*/
 #ifdef __NVCC__
 #pragma push
-#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
+#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
 #pragma nv_diag_suppress code_is_unreachable
 #else
 #pragma diag_suppress code_is_unreachable

Grid/algorithms/Algorithms.h

@@ -54,6 +54,7 @@ NAMESPACE_CHECK(BiCGSTAB);
 #include <Grid/algorithms/iterative/SchurRedBlack.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMultiShift.h>
 #include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
+#include <Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h>
 #include <Grid/algorithms/iterative/BiCGSTABMixedPrec.h>
 #include <Grid/algorithms/iterative/BlockConjugateGradient.h>
 #include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h>

Grid/algorithms/iterative/ConjugateGradientMixedPrecBatched.h

@@ -0,0 +1,213 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/ConjugateGradientMixedPrecBatched.h
+
+    Copyright (C) 2015
+
+    Author: Raoul Hodgson <raoul.hodgson@ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H
+#define GRID_CONJUGATE_GRADIENT_MIXED_PREC_BATCHED_H
+
+NAMESPACE_BEGIN(Grid);
+
+//Mixed precision restarted defect correction CG
+template<class FieldD,class FieldF, 
+  typename std::enable_if< getPrecision<FieldD>::value == 2, int>::type = 0,
+  typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0> 
+class MixedPrecisionConjugateGradientBatched : public LinearFunction<FieldD> {
+public:
+  using LinearFunction<FieldD>::operator();
+  RealD   Tolerance;
+  RealD   InnerTolerance; //Initial tolerance for inner CG. Defaults to Tolerance but can be changed
+  Integer MaxInnerIterations;
+  Integer MaxOuterIterations;
+  Integer MaxPatchupIterations;
+  GridBase* SinglePrecGrid; //Grid for single-precision fields
+  RealD OuterLoopNormMult; //Stop the outer loop and move to a final double prec solve when the residual is OuterLoopNormMult * Tolerance
+  LinearOperatorBase<FieldF> &Linop_f;
+  LinearOperatorBase<FieldD> &Linop_d;
+
+  //Option to speed up *inner single precision* solves using a LinearFunction that produces a guess
+  LinearFunction<FieldF> *guesser;
+  bool updateResidual;
+  
+  MixedPrecisionConjugateGradientBatched(RealD tol, 
+          Integer maxinnerit, 
+          Integer maxouterit, 
+          Integer maxpatchit,
+          GridBase* _sp_grid, 
+          LinearOperatorBase<FieldF> &_Linop_f, 
+          LinearOperatorBase<FieldD> &_Linop_d,
+          bool _updateResidual=true) :
+    Linop_f(_Linop_f), Linop_d(_Linop_d),
+    Tolerance(tol), InnerTolerance(tol), MaxInnerIterations(maxinnerit), MaxOuterIterations(maxouterit), MaxPatchupIterations(maxpatchit), SinglePrecGrid(_sp_grid),
+    OuterLoopNormMult(100.), guesser(NULL), updateResidual(_updateResidual) { };
+
+  void useGuesser(LinearFunction<FieldF> &g){
+    guesser = &g;
+  }
+  
+  void operator() (const FieldD &src_d_in, FieldD &sol_d){
+    std::vector<FieldD> srcs_d_in{src_d_in};
+    std::vector<FieldD> sols_d{sol_d};
+
+    (*this)(srcs_d_in,sols_d);
+
+    sol_d = sols_d[0];
+  }
+
+  void operator() (const std::vector<FieldD> &src_d_in, std::vector<FieldD> &sol_d){
+    assert(src_d_in.size() == sol_d.size());
+    int NBatch = src_d_in.size();
+
+    std::cout << GridLogMessage << "NBatch = " << NBatch << std::endl;
+
+    Integer TotalOuterIterations = 0; //Number of restarts
+    std::vector<Integer> TotalInnerIterations(NBatch,0);     //Number of inner CG iterations
+    std::vector<Integer> TotalFinalStepIterations(NBatch,0); //Number of CG iterations in final patch-up step
+  
+    GridStopWatch TotalTimer;
+    TotalTimer.Start();
+
+    GridStopWatch InnerCGtimer;
+    GridStopWatch PrecChangeTimer;
+    
+    int cb = src_d_in[0].Checkerboard();
+    
+    std::vector<RealD> src_norm;
+    std::vector<RealD> norm;
+    std::vector<RealD> stop;
+    
+    GridBase* DoublePrecGrid = src_d_in[0].Grid();
+    FieldD tmp_d(DoublePrecGrid);
+    tmp_d.Checkerboard() = cb;
+    
+    FieldD tmp2_d(DoublePrecGrid);
+    tmp2_d.Checkerboard() = cb;
+
+    std::vector<FieldD> src_d;
+    std::vector<FieldF> src_f;
+    std::vector<FieldF> sol_f;
+
+    for (int i=0; i<NBatch; i++) {
+      sol_d[i].Checkerboard() = cb;
+
+      src_norm.push_back(norm2(src_d_in[i]));
+      norm.push_back(0.);
+      stop.push_back(src_norm[i] * Tolerance*Tolerance);
+
+      src_d.push_back(src_d_in[i]); //source for next inner iteration, computed from residual during operation
+
+      src_f.push_back(SinglePrecGrid);
+      src_f[i].Checkerboard() = cb;
+
+      sol_f.push_back(SinglePrecGrid);
+      sol_f[i].Checkerboard() = cb;
+    }
+    
+    RealD inner_tol = InnerTolerance;
+    
+    ConjugateGradient<FieldF> CG_f(inner_tol, MaxInnerIterations);
+    CG_f.ErrorOnNoConverge = false;
+    
+    Integer &outer_iter = TotalOuterIterations; //so it will be equal to the final iteration count
+      
+    for(outer_iter = 0; outer_iter < MaxOuterIterations; outer_iter++){
+      std::cout << GridLogMessage << std::endl;
+      std::cout << GridLogMessage << "Outer iteration " << outer_iter << std::endl;
+      
+      bool allConverged = true;
+      
+      for (int i=0; i<NBatch; i++) {
+        //Compute double precision rsd and also new RHS vector.
+        Linop_d.HermOp(sol_d[i], tmp_d);
+        norm[i] = axpy_norm(src_d[i], -1., tmp_d, src_d_in[i]); //src_d is residual vector
+        
+        std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Outer iteration " << outer_iter <<" solve " << i << " residual "<< norm[i] << " target "<< stop[i] <<std::endl;
+
+        PrecChangeTimer.Start();
+        precisionChange(src_f[i], src_d[i]);
+        PrecChangeTimer.Stop();
+        
+        sol_f[i] = Zero();
+      
+        if(norm[i] > OuterLoopNormMult * stop[i]) {
+          allConverged = false;
+        }
+      }
+      if (allConverged) break;
+
+      if (updateResidual) {
+        RealD normMax = *std::max_element(std::begin(norm), std::end(norm));
+        RealD stopMax = *std::max_element(std::begin(stop), std::end(stop));
+        while( normMax * inner_tol * inner_tol < stopMax) inner_tol *= 2;  // inner_tol = sqrt(stop/norm) ??
+        CG_f.Tolerance = inner_tol;
+      }
+
+      //Optionally improve inner solver guess (eg using known eigenvectors)
+      if(guesser != NULL) {
+        (*guesser)(src_f, sol_f);
+      }
+
+      for (int i=0; i<NBatch; i++) {
+        //Inner CG
+        InnerCGtimer.Start();
+        CG_f(Linop_f, src_f[i], sol_f[i]);
+        InnerCGtimer.Stop();
+        TotalInnerIterations[i] += CG_f.IterationsToComplete;
+        
+        //Convert sol back to double and add to double prec solution
+        PrecChangeTimer.Start();
+        precisionChange(tmp_d, sol_f[i]);
+        PrecChangeTimer.Stop();
+        
+        axpy(sol_d[i], 1.0, tmp_d, sol_d[i]);
+      }
+
+    }
+    
+    //Final trial CG
+    std::cout << GridLogMessage << std::endl;
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Starting final patch-up double-precision solve"<<std::endl;
+    
+    for (int i=0; i<NBatch; i++) {
+      ConjugateGradient<FieldD> CG_d(Tolerance, MaxPatchupIterations);
+      CG_d(Linop_d, src_d_in[i], sol_d[i]);
+      TotalFinalStepIterations[i] += CG_d.IterationsToComplete;
+    }
+
+    TotalTimer.Stop();
+
+    std::cout << GridLogMessage << std::endl;
+    for (int i=0; i<NBatch; i++) {
+      std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: solve " << i << " Inner CG iterations " << TotalInnerIterations[i] << " Restarts " << TotalOuterIterations << " Final CG iterations " << TotalFinalStepIterations[i] << std::endl;
+    }
+    std::cout << GridLogMessage << std::endl;
+    std::cout<<GridLogMessage<<"MixedPrecisionConjugateGradientBatched: Total time " << TotalTimer.Elapsed() << " Precision change " << PrecChangeTimer.Elapsed() << " Inner CG total " << InnerCGtimer.Elapsed() << std::endl;
+    
+  }
+};
+
+NAMESPACE_END(Grid);
+
+#endif

Grid/allocator/MemoryManager.cc

@@ -4,11 +4,14 @@ NAMESPACE_BEGIN(Grid);
 
 /*Allocation types, saying which pointer cache should be used*/
 #define Cpu      (0)
-#define CpuSmall (1)
+#define CpuHuge  (1)
-#define Acc      (2)
+#define CpuSmall (2)
-#define AccSmall (3)
+#define Acc      (3)
-#define Shared   (4)
+#define AccHuge  (4)
-#define SharedSmall (5)
+#define AccSmall (5)
+#define Shared   (6)
+#define SharedHuge  (7)
+#define SharedSmall (8)
 #undef GRID_MM_VERBOSE 
 uint64_t total_shared;
 uint64_t total_device;
@@ -35,12 +38,15 @@ void MemoryManager::PrintBytes(void)
   
 }
 
+uint64_t MemoryManager::DeviceCacheBytes() { return CacheBytes[Acc] + CacheBytes[AccHuge] + CacheBytes[AccSmall]; }
+uint64_t MemoryManager::HostCacheBytes()   { return CacheBytes[Cpu] + CacheBytes[CpuHuge] + CacheBytes[CpuSmall]; }
+
 //////////////////////////////////////////////////////////////////////
 // Data tables for recently freed pooiniter caches
 //////////////////////////////////////////////////////////////////////
 MemoryManager::AllocationCacheEntry MemoryManager::Entries[MemoryManager::NallocType][MemoryManager::NallocCacheMax];
 int MemoryManager::Victim[MemoryManager::NallocType];
-int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 8, 2, 8, 2, 8 };
+int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 0, 8, 8, 0, 16, 8, 0, 16 };
 uint64_t MemoryManager::CacheBytes[MemoryManager::NallocType];
 //////////////////////////////////////////////////////////////////////
 // Actual allocation and deallocation utils
@@ -170,6 +176,16 @@ void MemoryManager::Init(void)
     }
   }
 
+  str= getenv("GRID_ALLOC_NCACHE_HUGE");
+  if ( str ) {
+    Nc = atoi(str);
+    if ( (Nc>=0) && (Nc < NallocCacheMax)) {
+      Ncache[CpuHuge]=Nc;
+      Ncache[AccHuge]=Nc;
+      Ncache[SharedHuge]=Nc;
+    }
+  }
+
   str= getenv("GRID_ALLOC_NCACHE_SMALL");
   if ( str ) {
     Nc = atoi(str);
@@ -190,7 +206,9 @@ void MemoryManager::InitMessage(void) {
   
   std::cout << GridLogMessage<< "MemoryManager::Init() setting up"<<std::endl;
 #ifdef ALLOCATION_CACHE
-  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<std::endl;
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent host   allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<" HUGE "<<Ncache[CpuHuge]<<std::endl;
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent device allocations: SMALL "<<Ncache[AccSmall]<<" LARGE "<<Ncache[Acc]<<" Huge "<<Ncache[AccHuge]<<std::endl;
+  std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent shared allocations: SMALL "<<Ncache[SharedSmall]<<" LARGE "<<Ncache[Shared]<<" Huge "<<Ncache[SharedHuge]<<std::endl;
 #endif
   
 #ifdef GRID_UVM
@@ -222,8 +240,11 @@ void MemoryManager::InitMessage(void) {
 void *MemoryManager::Insert(void *ptr,size_t bytes,int type) 
 {
 #ifdef ALLOCATION_CACHE
-  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
+  int cache;
-  int cache = type + small;
+  if      (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2;
+  else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1;
+  else                                     cache = type;
+
   return Insert(ptr,bytes,Entries[cache],Ncache[cache],Victim[cache],CacheBytes[cache]);  
 #else
   return ptr;
@@ -232,11 +253,12 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,int type)
 
 void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim, uint64_t &cacheBytes) 
 {
-  assert(ncache>0);
 #ifdef GRID_OMP
   assert(omp_in_parallel()==0);
 #endif 
 
+  if (ncache == 0) return ptr;
+
   void * ret = NULL;
   int v = -1;
 
@@ -271,8 +293,11 @@ void *MemoryManager::Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries
 void *MemoryManager::Lookup(size_t bytes,int type)
 {
 #ifdef ALLOCATION_CACHE
-  bool small = (bytes < GRID_ALLOC_SMALL_LIMIT);
+  int cache;
-  int cache = type+small;
+  if      (bytes < GRID_ALLOC_SMALL_LIMIT) cache = type + 2;
+  else if (bytes >= GRID_ALLOC_HUGE_LIMIT) cache = type + 1;
+  else                                     cache = type;
+
   return Lookup(bytes,Entries[cache],Ncache[cache],CacheBytes[cache]);
 #else
   return NULL;
@@ -281,7 +306,6 @@ void *MemoryManager::Lookup(size_t bytes,int type)
 
 void *MemoryManager::Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t & cacheBytes) 
 {
-  assert(ncache>0);
 #ifdef GRID_OMP
   assert(omp_in_parallel()==0);
 #endif 

Grid/allocator/MemoryManager.h

@@ -35,6 +35,12 @@ NAMESPACE_BEGIN(Grid);
 // Move control to configure.ac and Config.h?
 
 #define GRID_ALLOC_SMALL_LIMIT (4096)
+#define GRID_ALLOC_HUGE_LIMIT  (2147483648)
+
+#define STRINGIFY(x) #x
+#define TOSTRING(x) STRINGIFY(x)
+#define FILE_LINE __FILE__ ":" TOSTRING(__LINE__)
+#define AUDIT(a) MemoryManager::Audit(FILE_LINE)
 
 /*Pinning pages is costly*/
 ////////////////////////////////////////////////////////////////////////////
@@ -65,6 +71,21 @@ enum ViewMode {
   CpuWriteDiscard = 0x10 // same for now
 };
 
+struct MemoryStatus {
+  uint64_t     DeviceBytes;
+  uint64_t     DeviceLRUBytes;
+  uint64_t     DeviceMaxBytes;
+  uint64_t     HostToDeviceBytes;
+  uint64_t     DeviceToHostBytes;
+  uint64_t     HostToDeviceXfer;
+  uint64_t     DeviceToHostXfer;
+  uint64_t     DeviceEvictions;
+  uint64_t     DeviceDestroy;
+  uint64_t     DeviceAllocCacheBytes;
+  uint64_t     HostAllocCacheBytes;
+};
+
+
 class MemoryManager {
 private:
 
@@ -78,7 +99,7 @@ private:
   } AllocationCacheEntry;
 
   static const int NallocCacheMax=128; 
-  static const int NallocType=6;
+  static const int NallocType=9;
   static AllocationCacheEntry Entries[NallocType][NallocCacheMax];
   static int Victim[NallocType];
   static int Ncache[NallocType];
@@ -92,8 +113,9 @@ private:
   static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim,uint64_t &cbytes) ;
   static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t &cbytes) ;
 
-  static void PrintBytes(void);
  public:
+  static void PrintBytes(void);
+  static void Audit(std::string s);
   static void Init(void);
   static void InitMessage(void);
   static void *AcceleratorAllocate(size_t bytes);
@@ -113,6 +135,27 @@ private:
   static uint64_t     DeviceToHostBytes;
   static uint64_t     HostToDeviceXfer;
   static uint64_t     DeviceToHostXfer;
+  static uint64_t     DeviceEvictions;
+  static uint64_t     DeviceDestroy;
+  
+  static uint64_t     DeviceCacheBytes();
+  static uint64_t     HostCacheBytes();
+
+  static MemoryStatus GetFootprint(void) {
+    MemoryStatus stat;
+    stat.DeviceBytes       = DeviceBytes;
+    stat.DeviceLRUBytes    = DeviceLRUBytes;
+    stat.DeviceMaxBytes    = DeviceMaxBytes;
+    stat.HostToDeviceBytes = HostToDeviceBytes;
+    stat.DeviceToHostBytes = DeviceToHostBytes;
+    stat.HostToDeviceXfer  = HostToDeviceXfer;
+    stat.DeviceToHostXfer  = DeviceToHostXfer;
+    stat.DeviceEvictions   = DeviceEvictions;
+    stat.DeviceDestroy     = DeviceDestroy;
+    stat.DeviceAllocCacheBytes = DeviceCacheBytes();
+    stat.HostAllocCacheBytes   = HostCacheBytes();
+    return stat;
+  };
   
  private:
 #ifndef GRID_UVM
@@ -170,6 +213,7 @@ private:
 
  public:
   static void Print(void);
+  static void PrintAll(void);
   static void PrintState( void* CpuPtr);
   static int   isOpen   (void* CpuPtr);
   static void  ViewClose(void* CpuPtr,ViewMode mode);

Grid/allocator/MemoryManagerCache.cc

@@ -3,8 +3,13 @@
 
 #warning "Using explicit device memory copies"
 NAMESPACE_BEGIN(Grid);
-//#define dprintf(...) printf ( __VA_ARGS__ ); fflush(stdout);
+
-#define dprintf(...)
+#define MAXLINE 512
+static char print_buffer [ MAXLINE ];
+
+#define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
+#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
+//#define dprintf(...) 
 
 
 ////////////////////////////////////////////////////////////
@@ -23,6 +28,8 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
+uint64_t  MemoryManager::DeviceEvictions;
+uint64_t  MemoryManager::DeviceDestroy;
 
 ////////////////////////////////////
 // Priority ordering for unlocked entries
@@ -104,15 +111,17 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
   ///////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-   dprintf("MemoryManager: Discard(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  mprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
   assert(AccCache.accLock==0);
   assert(AccCache.cpuLock==0);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
+    DeviceDestroy++;
     DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
-    dprintf("MemoryManager: Free(%llx) LRU %lld Total %lld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
+    AccCache.AccPtr=(uint64_t) NULL;
+    dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);  
   }
   uint64_t CpuPtr = AccCache.CpuPtr;
   EntryErase(CpuPtr);
@@ -121,26 +130,36 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
 {
   ///////////////////////////////////////////////////////////////////////////
-  // Make CPU consistent, remove from Accelerator, remove entry
+  // Make CPU consistent, remove from Accelerator, remove from LRU, LEAVE CPU only entry
-  // Cannot be locked. If allocated must be in LRU pool.
+  // Cannot be acclocked. If allocated must be in LRU pool.
+  //
+  // Nov 2022... Felix issue: Allocating two CpuPtrs, can have an entry in LRU-q with CPUlock.
+  //                          and require to evict the AccPtr copy. Eviction was a mistake in CpuViewOpen
+  //                          but there is a weakness where CpuLock entries are attempted for erase
+  //                          Take these OUT LRU queue when CPU locked?
+  //                          Cannot take out the table as cpuLock data is important.
   ///////////////////////////////////////////////////////////////////////////
   assert(AccCache.state!=Empty);
   
-  dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr); 
+  mprintf("MemoryManager: Evict cpu %lx acc %lx cpuLock %ld accLock %ld\n",
-  assert(AccCache.accLock==0);
+	  (uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
-  assert(AccCache.cpuLock==0);
+	  (uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock); 
+  assert(AccCache.accLock==0); // Cannot evict so logic bomb
+  assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.state==AccDirty) {
     Flush(AccCache);
   }
-  assert(AccCache.CpuPtr!=(uint64_t)NULL);
   if(AccCache.AccPtr) {
     AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
-    DeviceBytes   -=AccCache.bytes;
     LRUremove(AccCache);
-    dprintf("MemoryManager: Free(%llx) footprint now %lld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
+    AccCache.AccPtr=(uint64_t)NULL;
+    AccCache.state=CpuDirty; // CPU primary now
+    DeviceBytes   -=AccCache.bytes;
+    dprintf("MemoryManager: Free(%lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);  
   }
-  uint64_t CpuPtr = AccCache.CpuPtr;
+  //  uint64_t CpuPtr = AccCache.CpuPtr;
-  EntryErase(CpuPtr);
+  DeviceEvictions++;
+  //  EntryErase(CpuPtr);
 }
 void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
 {
@@ -150,7 +169,7 @@ void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
   assert(AccCache.AccPtr!=(uint64_t)NULL);
   assert(AccCache.CpuPtr!=(uint64_t)NULL);
   acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes);
-  dprintf("MemoryManager: Flush  %llx -> %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: Flush  %lx -> %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   DeviceToHostBytes+=AccCache.bytes;
   DeviceToHostXfer++;
   AccCache.state=Consistent;
@@ -165,7 +184,7 @@ void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
     AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
     DeviceBytes+=AccCache.bytes;
   }
-  dprintf("MemoryManager: Clone %llx <- %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
+  mprintf("MemoryManager: Clone %lx <- %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
   acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
   HostToDeviceBytes+=AccCache.bytes;
   HostToDeviceXfer++;
@@ -191,6 +210,7 @@ void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
 void MemoryManager::ViewClose(void* Ptr,ViewMode mode)
 {
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
+    dprintf("AcceleratorViewClose %lx\n",(uint64_t)Ptr);
     AcceleratorViewClose((uint64_t)Ptr);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     CpuViewClose((uint64_t)Ptr);
@@ -202,6 +222,7 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
 {
   uint64_t CpuPtr = (uint64_t)_CpuPtr;
   if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
+    dprintf("AcceleratorViewOpen %lx\n",(uint64_t)CpuPtr);
     return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint);
   } else if( (mode==CpuRead)||(mode==CpuWrite)){
     return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint);
@@ -212,13 +233,16 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
 }
 void  MemoryManager::EvictVictims(uint64_t bytes)
 {
+  assert(bytes<DeviceMaxBytes);
   while(bytes+DeviceLRUBytes > DeviceMaxBytes){
     if ( DeviceLRUBytes > 0){
       assert(LRU.size()>0);
-      uint64_t victim = LRU.back();
+      uint64_t victim = LRU.back(); // From the LRU
       auto AccCacheIterator = EntryLookup(victim);
       auto & AccCache = AccCacheIterator->second;
       Evict(AccCache);
+    } else {
+      return;
     }
   }
 }
@@ -241,11 +265,12 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
   assert(AccCache.cpuLock==0);  // Programming error
 
   if(AccCache.state!=Empty) {
-    dprintf("ViewOpen found entry %llx %llx : %lld %lld\n",
+    dprintf("ViewOpen found entry %lx %lx : %ld %ld accLock %ld\n",
 		    (uint64_t)AccCache.CpuPtr,
 		    (uint64_t)CpuPtr,
 		    (uint64_t)AccCache.bytes,
-		    (uint64_t)bytes);
+	            (uint64_t)bytes,
+		    (uint64_t)AccCache.accLock);
     assert(AccCache.CpuPtr == CpuPtr);
     assert(AccCache.bytes  ==bytes);
   }
@@ -280,6 +305,7 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // Empty + AccRead => Consistent
     }
     AccCache.accLock= 1;
+    dprintf("Copied Empty entry into device accLock= %d\n",AccCache.accLock);
   } else if(AccCache.state==CpuDirty ){
     if(mode==AcceleratorWriteDiscard) {
       CpuDiscard(AccCache);
@@ -292,28 +318,30 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
       AccCache.state  = Consistent; // CpuDirty + AccRead => Consistent
     }
     AccCache.accLock++;
-    dprintf("Copied CpuDirty entry into device accLock %d\n",AccCache.accLock);
+    dprintf("CpuDirty entry into device ++accLock= %d\n",AccCache.accLock);
   } else if(AccCache.state==Consistent) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty;   // Consistent + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = Consistent; // Consistent + AccRead => Consistent
     AccCache.accLock++;
-    dprintf("Consistent entry into device accLock %d\n",AccCache.accLock);
+    dprintf("Consistent entry into device ++accLock= %d\n",AccCache.accLock);
   } else if(AccCache.state==AccDirty) {
     if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
       AccCache.state  = AccDirty; // AccDirty + AcceleratorWrite=> AccDirty
     else
       AccCache.state  = AccDirty; // AccDirty + AccRead => AccDirty
     AccCache.accLock++;
-    dprintf("AccDirty entry into device accLock %d\n",AccCache.accLock);
+    dprintf("AccDirty entry ++accLock= %d\n",AccCache.accLock);
   } else {
     assert(0);
   }
 
-  // If view is opened on device remove from LRU
+  assert(AccCache.accLock>0);
+  // If view is opened on device must remove from LRU
   if(AccCache.LRU_valid==1){
     // must possibly remove from LRU as now locked on GPU
+    dprintf("AccCache entry removed from LRU \n");
     LRUremove(AccCache);
   }
 
@@ -334,10 +362,12 @@ void MemoryManager::AcceleratorViewClose(uint64_t CpuPtr)
   assert(AccCache.accLock>0);
 
   AccCache.accLock--;
-
   // Move to LRU queue if not locked and close on device
   if(AccCache.accLock==0) {
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
     LRUinsert(AccCache);
+  } else {
+    dprintf("AccleratorViewClose %lx AccLock decremented to %ld\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
   }
 }
 void MemoryManager::CpuViewClose(uint64_t CpuPtr)
@@ -374,9 +404,10 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
   auto AccCacheIterator = EntryLookup(CpuPtr);
   auto & AccCache = AccCacheIterator->second;
 
-  if (!AccCache.AccPtr) {
+  // CPU doesn't need to free space
-     EvictVictims(bytes);
+  //  if (!AccCache.AccPtr) {
-  }
+  //    EvictVictims(bytes);
+  //  }
 
   assert((mode==CpuRead)||(mode==CpuWrite));
   assert(AccCache.accLock==0);  // Programming error
@@ -430,20 +461,28 @@ void  MemoryManager::NotifyDeletion(void *_ptr)
 void  MemoryManager::Print(void)
 {
   PrintBytes();
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
-  std::cout << GridLogDebug << "Memory Manager                             " << std::endl;
+  std::cout << GridLogMessage << "Memory Manager                             " << std::endl;
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
-  std::cout << GridLogDebug << DeviceBytes   << " bytes allocated on device " << std::endl;
+  std::cout << GridLogMessage << DeviceBytes   << " bytes allocated on device " << std::endl;
-  std::cout << GridLogDebug << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
+  std::cout << GridLogMessage << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
-  std::cout << GridLogDebug << DeviceMaxBytes<< " bytes max on device       " << std::endl;
+  std::cout << GridLogMessage << DeviceMaxBytes<< " bytes max on device       " << std::endl;
-  std::cout << GridLogDebug << HostToDeviceXfer << " transfers        to   device " << std::endl;
+  std::cout << GridLogMessage << HostToDeviceXfer << " transfers        to   device " << std::endl;
-  std::cout << GridLogDebug << DeviceToHostXfer << " transfers        from device " << std::endl;
+  std::cout << GridLogMessage << DeviceToHostXfer << " transfers        from device " << std::endl;
-  std::cout << GridLogDebug << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
+  std::cout << GridLogMessage << HostToDeviceBytes<< " bytes transfered to   device " << std::endl;
-  std::cout << GridLogDebug << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
+  std::cout << GridLogMessage << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
-  std::cout << GridLogDebug << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
+  std::cout << GridLogMessage << DeviceEvictions  << " Evictions from device " << std::endl;
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << DeviceDestroy    << " Destroyed vectors on device " << std::endl;
-  std::cout << GridLogDebug << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+  std::cout << GridLogMessage << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+}
+void  MemoryManager::PrintAll(void)
+{
+  Print();
+  std::cout << GridLogMessage << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
   for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
     auto &AccCache = it->second;
     
@@ -453,13 +492,13 @@ void  MemoryManager::Print(void)
     if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
     if ( AccCache.state==Consistent)str = std::string("Consistent");
 
-    std::cout << GridLogDebug << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
 	      << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
 	      << "\t" << AccCache.cpuLock
 	      << "\t" << AccCache.accLock
 	      << "\t" << AccCache.LRU_valid<<std::endl;
   }
-  std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
+  std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
 
 };
 int   MemoryManager::isOpen   (void* _CpuPtr) 
@@ -473,6 +512,61 @@ int   MemoryManager::isOpen   (void* _CpuPtr)
     return 0;
   }
 }
+void MemoryManager::Audit(std::string s)
+{
+  uint64_t CpuBytes=0;
+  uint64_t AccBytes=0;
+  uint64_t LruBytes1=0;
+  uint64_t LruBytes2=0;
+  uint64_t LruCnt=0;
+  uint64_t LockedBytes=0;
+  
+  std::cout << " Memory Manager::Audit() from "<<s<<std::endl;
+  for(auto it=LRU.begin();it!=LRU.end();it++){
+    uint64_t cpuPtr = *it;
+    assert(EntryPresent(cpuPtr));
+    auto AccCacheIterator = EntryLookup(cpuPtr);
+    auto & AccCache = AccCacheIterator->second;
+    LruBytes2+=AccCache.bytes;
+    assert(AccCache.LRU_valid==1);
+    assert(AccCache.LRU_entry==it);
+  }
+  std::cout << " Memory Manager::Audit() LRU queue matches table entries "<<std::endl;
+  for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
+    auto &AccCache = it->second;
+    
+    std::string str;
+    if ( AccCache.state==Empty    ) str = std::string("Empty");
+    if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
+    if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
+    if ( AccCache.state==Consistent)str = std::string("Consistent");
+
+    CpuBytes+=AccCache.bytes;
+    if( AccCache.AccPtr )    AccBytes+=AccCache.bytes;
+    if( AccCache.LRU_valid ) LruBytes1+=AccCache.bytes;
+    if( AccCache.LRU_valid ) LruCnt++;
+    
+    if ( AccCache.cpuLock || AccCache.accLock ) {
+      assert(AccCache.LRU_valid==0);
+      std::cout << GridLogError << s<< "\n\t 0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+		<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+		<< "\t cpuLock  " << AccCache.cpuLock
+		<< "\t accLock  " << AccCache.accLock
+		<< "\t LRUvalid " << AccCache.LRU_valid<<std::endl;
+    }
+
+    assert( AccCache.cpuLock== 0 ) ;
+    assert( AccCache.accLock== 0 ) ;
+  }
+  std::cout << " Memory Manager::Audit() no locked table entries "<<std::endl;
+  assert(LruBytes1==LruBytes2);
+  assert(LruBytes1==DeviceLRUBytes);
+  std::cout << " Memory Manager::Audit() evictable bytes matches sum over table "<<std::endl;
+  assert(AccBytes==DeviceBytes);
+  std::cout << " Memory Manager::Audit() device bytes matches sum over table "<<std::endl;
+  assert(LruCnt == LRU.size());
+  std::cout << " Memory Manager::Audit() LRU entry count matches "<<std::endl;
+}
 
 void MemoryManager::PrintState(void* _CpuPtr)
 {
@@ -489,8 +583,8 @@ void MemoryManager::PrintState(void* _CpuPtr)
     if ( AccCache.state==EvictNext) str = std::string("EvictNext");
 
     std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
-    std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
+    std::cout << GridLogMessage << "\tx"<<std::hex<<AccCache.CpuPtr<<std::dec
-    << "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
+    << "\tx"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
     << "\t" << AccCache.cpuLock
     << "\t" << AccCache.accLock
     << "\t" << AccCache.LRU_valid<<std::endl;

Grid/allocator/MemoryManagerShared.cc

@@ -12,7 +12,10 @@ uint64_t  MemoryManager::HostToDeviceBytes;
 uint64_t  MemoryManager::DeviceToHostBytes;
 uint64_t  MemoryManager::HostToDeviceXfer;
 uint64_t  MemoryManager::DeviceToHostXfer;
+uint64_t  MemoryManager::DeviceEvictions;
+uint64_t  MemoryManager::DeviceDestroy;
 
+void  MemoryManager::Audit(std::string s){};
 void  MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
 void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };
 int   MemoryManager::isOpen   (void* CpuPtr) { return 0;}
@@ -21,6 +24,7 @@ void  MemoryManager::PrintState(void* CpuPtr)
 std::cout << GridLogMessage << "Host<->Device memory movement not currently managed by Grid." << std::endl;
 };
 void  MemoryManager::Print(void){};
+void  MemoryManager::PrintAll(void){};
 void  MemoryManager::NotifyDeletion(void *ptr){};
 
 NAMESPACE_END(Grid);

Grid/communicator/Communicator_mpi3.cc

@@ -401,8 +401,6 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
 {
   //   std::cout << "Copy Synchronised\n"<<std::endl;
-  acceleratorCopySynchronise();
-
   int nreq=list.size();
 
   if (nreq==0) return;

Grid/communicator/SharedMemoryMPI.cc

@@ -36,10 +36,11 @@ Author: Christoph Lehner <christoph@lhnr.de>
 #ifdef GRID_HIP
 #include <hip/hip_runtime_api.h>
 #endif
-#ifdef GRID_SYCl
+#ifdef GRID_SYCL
-
+#define GRID_SYCL_LEVEL_ZERO_IPC
 #endif
 
+
 NAMESPACE_BEGIN(Grid); 
 #define header "SharedMemoryMpi: "
 /*Construct from an MPI communicator*/

Grid/cshift/Cshift_common.h

@@ -297,6 +297,30 @@ template<class vobj> void Scatter_plane_merge(Lattice<vobj> &rhs,ExtractPointerA
   }
 }
 
+#if (defined(GRID_CUDA) || defined(GRID_HIP)) && defined(ACCELERATOR_CSHIFT)
+
+template <typename T>
+T iDivUp(T a, T b) // Round a / b to nearest higher integer value
+{ return (a % b != 0) ? (a / b + 1) : (a / b); }
+
+template <typename T>
+__global__ void populate_Cshift_table(T* vector, T lo, T ro, T e1, T e2, T stride)
+{
+    int idx = blockIdx.x*blockDim.x + threadIdx.x;
+    if (idx >= e1*e2) return;
+
+    int n, b, o;
+
+    n = idx / e2;
+    b = idx % e2;
+    o = n*stride + b;
+
+    vector[2*idx + 0] = lo + o;
+    vector[2*idx + 1] = ro + o;
+}
+
+#endif
+
 //////////////////////////////////////////////////////
 // local to node block strided copies
 //////////////////////////////////////////////////////
@@ -321,12 +345,20 @@ template<class vobj> void Copy_plane(Lattice<vobj>& lhs,const Lattice<vobj> &rhs
   int ent=0;
 
   if(cbmask == 0x3 ){
+#if (defined(GRID_CUDA) || defined(GRID_HIP)) && defined(ACCELERATOR_CSHIFT)
+    ent = e1*e2;
+    dim3 blockSize(acceleratorThreads());
+    dim3 gridSize(iDivUp((unsigned int)ent, blockSize.x));
+    populate_Cshift_table<<<gridSize, blockSize>>>(&Cshift_table[0].first, lo, ro, e1, e2, stride);
+    accelerator_barrier();
+#else
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
         int o =n*stride+b;
 	Cshift_table[ent++] = std::pair<int,int>(lo+o,ro+o);
       }
     }
+#endif
   } else { 
     for(int n=0;n<e1;n++){
       for(int b=0;b<e2;b++){
@@ -377,11 +409,19 @@ template<class vobj> void Copy_plane_permute(Lattice<vobj>& lhs,const Lattice<vo
   int ent=0;
 
   if ( cbmask == 0x3 ) {
+#if (defined(GRID_CUDA) || defined(GRID_HIP)) && defined(ACCELERATOR_CSHIFT)
+    ent = e1*e2;
+    dim3 blockSize(acceleratorThreads());
+    dim3 gridSize(iDivUp((unsigned int)ent, blockSize.x));
+    populate_Cshift_table<<<gridSize, blockSize>>>(&Cshift_table[0].first, lo, ro, e1, e2, stride);
+    accelerator_barrier();
+#else
     for(int n=0;n<e1;n++){
     for(int b=0;b<e2;b++){
       int o  =n*stride;
       Cshift_table[ent++] = std::pair<int,int>(lo+o+b,ro+o+b);
     }}
+#endif
   } else {
     for(int n=0;n<e1;n++){
     for(int b=0;b<e2;b++){

Grid/lattice/Lattice.h

@@ -46,3 +46,4 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/lattice/Lattice_unary.h>
 #include <Grid/lattice/Lattice_transfer.h>
 #include <Grid/lattice/Lattice_basis.h>
+#include <Grid/lattice/Lattice_crc.h>

Grid/lattice/Lattice_base.h

@@ -288,8 +288,8 @@ public:
     typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
     conformable(*this,r);
     this->checkerboard = r.Checkerboard();
-    auto me =   View(AcceleratorWriteDiscard);
     auto him= r.View(AcceleratorRead);
+    auto me =   View(AcceleratorWriteDiscard);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });
@@ -303,8 +303,8 @@ public:
   inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
     this->checkerboard = r.Checkerboard();
     conformable(*this,r);
-    auto me =   View(AcceleratorWriteDiscard);
     auto him= r.View(AcceleratorRead);
+    auto me =   View(AcceleratorWriteDiscard);
     accelerator_for(ss,me.size(),vobj::Nsimd(),{
       coalescedWrite(me[ss],him(ss));
     });

Grid/lattice/Lattice_crc.h

@@ -0,0 +1,55 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/lattice/Lattice_crc.h
+
+    Copyright (C) 2021
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+template<class vobj> void DumpSliceNorm(std::string s,Lattice<vobj> &f,int mu=-1)
+{
+  auto ff = localNorm2(f);
+  if ( mu==-1 ) mu = f.Grid()->Nd()-1;
+  typedef typename vobj::tensor_reduced normtype;
+  typedef typename normtype::scalar_object scalar;
+  std::vector<scalar> sff;
+  sliceSum(ff,sff,mu);
+  for(int t=0;t<sff.size();t++){
+    std::cout << s<<" "<<t<<" "<<sff[t]<<std::endl;
+  }
+}
+
+template<class vobj> uint32_t crc(Lattice<vobj> & buf)
+{
+  autoView( buf_v , buf, CpuRead);
+  return ::crc32(0L,(unsigned char *)&buf_v[0],(size_t)sizeof(vobj)*buf.oSites());
+}
+
+#define CRC(U) std::cout << "FingerPrint "<<__FILE__ <<" "<< __LINE__ <<" "<< #U <<" "<<crc(U)<<std::endl;
+
+NAMESPACE_END(Grid);
+
+

Grid/lattice/Lattice_reduction.h

@@ -156,33 +156,44 @@ inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
 }
 
 template<class vobj>
-inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
+inline typename vobj::scalar_object rankSum(const Lattice<vobj> &arg)
 {
   Integer osites = arg.Grid()->oSites();
 #if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
-  typename vobj::scalar_object ssum;
   autoView( arg_v, arg, AcceleratorRead);
-  ssum= sum_gpu(&arg_v[0],osites);
+  return sum_gpu(&arg_v[0],osites);
 #else
   autoView(arg_v, arg, CpuRead);
-  auto ssum= sum_cpu(&arg_v[0],osites);
+  return sum_cpu(&arg_v[0],osites);
 #endif  
+}
+
+template<class vobj>
+inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
+{
+  auto ssum = rankSum(arg);
   arg.Grid()->GlobalSum(ssum);
   return ssum;
 }
 
 template<class vobj>
-inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
+inline typename vobj::scalar_object rankSumLarge(const Lattice<vobj> &arg)
 {
 #if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
   autoView( arg_v, arg, AcceleratorRead);
   Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_gpu_large(&arg_v[0],osites);
+  return sum_gpu_large(&arg_v[0],osites);
 #else
   autoView(arg_v, arg, CpuRead);
   Integer osites = arg.Grid()->oSites();
-  auto ssum= sum_cpu(&arg_v[0],osites);
+  return sum_cpu(&arg_v[0],osites);
 #endif
+}
+
+template<class vobj>
+inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
+{
+  auto ssum = rankSumLarge(arg);
   arg.Grid()->GlobalSum(ssum);
   return ssum;
 }

Grid/lattice/Lattice_transfer.h

@@ -288,7 +288,36 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
     blockZAXPY(fineDataRed,ip,Basis[v],fineDataRed); 
   }
 }
+template<class vobj,class CComplex,int nbasis,class VLattice>
+inline void batchBlockProject(std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
+                               const std::vector<Lattice<vobj>> &fineData,
+                               const VLattice &Basis)
+{
+  int NBatch = fineData.size();
+  assert(coarseData.size() == NBatch);
 
+  GridBase * fine  = fineData[0].Grid();
+  GridBase * coarse= coarseData[0].Grid();
+
+  Lattice<iScalar<CComplex>> ip(coarse);
+  std::vector<Lattice<vobj>> fineDataCopy = fineData;
+
+  autoView(ip_, ip, AcceleratorWrite);
+  for(int v=0;v<nbasis;v++) {
+    for (int k=0; k<NBatch; k++) {
+      autoView( coarseData_ , coarseData[k], AcceleratorWrite);
+      blockInnerProductD(ip,Basis[v],fineDataCopy[k]); // ip = <basis|fine>
+      accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
+        convertType(coarseData_[sc](v),ip_[sc]);
+      });
+
+      // improve numerical stability of projection
+      // |fine> = |fine> - <basis|fine> |basis>
+      ip=-ip;
+      blockZAXPY(fineDataCopy[k],ip,Basis[v],fineDataCopy[k]); 
+    }
+  }
+}
 
 template<class vobj,class vobj2,class CComplex>
   inline void blockZAXPY(Lattice<vobj> &fineZ,
@@ -590,6 +619,26 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
 }
 #endif
 
+template<class vobj,class CComplex,int nbasis,class VLattice>
+inline void batchBlockPromote(const std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
+                               std::vector<Lattice<vobj>> &fineData,
+                               const VLattice &Basis)
+{
+  int NBatch = coarseData.size();
+  assert(fineData.size() == NBatch);
+
+  GridBase * fine   = fineData[0].Grid();
+  GridBase * coarse = coarseData[0].Grid();
+  for (int k=0; k<NBatch; k++)
+    fineData[k]=Zero();
+  for (int i=0;i<nbasis;i++) {
+    for (int k=0; k<NBatch; k++) {
+      Lattice<iScalar<CComplex>> ip = PeekIndex<0>(coarseData[k],i);
+      blockZAXPY(fineData[k],ip,Basis[i],fineData[k]);
+    }
+  }
+}
+
 // Useful for precision conversion, or indeed anything where an operator= does a conversion on scalars.
 // Simd layouts need not match since we use peek/poke Local
 template<class vobj,class vvobj>

Grid/log/Log.cc

@@ -65,29 +65,40 @@ GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
 GridLogger GridLogError  (1, "Error" , GridLogColours, "RED");
 GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW");
 GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL");
+GridLogger GridLogMemory (1, "Memory", GridLogColours, "NORMAL");
+GridLogger GridLogTracing(1, "Tracing", GridLogColours, "NORMAL");
 GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
+GridLogger GridLogDslash     (1, "Dslash", GridLogColours, "BLUE");
 GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE");
 GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE");
+GridLogger GridLogHMC (1, "HMC", GridLogColours, "BLUE");
 
 void GridLogConfigure(std::vector<std::string> &logstreams) {
-  GridLogError.Active(0);
+  GridLogError.Active(1);
   GridLogWarning.Active(0);
   GridLogMessage.Active(1); // at least the messages should be always on
+  GridLogMemory.Active(0); 
+  GridLogTracing.Active(0); 
   GridLogIterative.Active(0);
   GridLogDebug.Active(0);
   GridLogPerformance.Active(0);
+  GridLogDslash.Active(0);
   GridLogIntegrator.Active(1);
   GridLogColours.Active(0);
+  GridLogHMC.Active(1);
 
   for (int i = 0; i < logstreams.size(); i++) {
-    if (logstreams[i] == std::string("Error"))       GridLogError.Active(1);
+    if (logstreams[i] == std::string("Tracing"))     GridLogTracing.Active(1);
+    if (logstreams[i] == std::string("Memory"))      GridLogMemory.Active(1);
     if (logstreams[i] == std::string("Warning"))     GridLogWarning.Active(1);
     if (logstreams[i] == std::string("NoMessage"))   GridLogMessage.Active(0);
     if (logstreams[i] == std::string("Iterative"))   GridLogIterative.Active(1);
     if (logstreams[i] == std::string("Debug"))       GridLogDebug.Active(1);
     if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1);
-    if (logstreams[i] == std::string("Integrator"))  GridLogIntegrator.Active(1);
+    if (logstreams[i] == std::string("Dslash"))      GridLogDslash.Active(1);
+    if (logstreams[i] == std::string("NoIntegrator"))GridLogIntegrator.Active(0);
+    if (logstreams[i] == std::string("NoHMC"))       GridLogHMC.Active(0);
     if (logstreams[i] == std::string("Colours"))     GridLogColours.Active(1);
   }
 }

Grid/log/Log.h

@@ -138,7 +138,8 @@ public:
         stream << std::setw(log.topWidth);
       }
       stream << log.topName << log.background()<< " : ";
-      stream << log.colour() <<  std::left;
+      //      stream << log.colour() <<  std::left;
+      stream <<  std::left;
       if (log.chanWidth > 0)
       {
         stream << std::setw(log.chanWidth);
@@ -153,9 +154,9 @@ public:
 	stream << log.evidence()
 	       << now	       << log.background() << " : " ;
       }
-      stream << log.colour();
+      //      stream << log.colour();
+      stream <<  std::right;
       stream.flags(f);
-
       return stream;
     } else { 
       return devnull;
@@ -180,8 +181,12 @@ extern GridLogger GridLogWarning;
 extern GridLogger GridLogMessage;
 extern GridLogger GridLogDebug  ;
 extern GridLogger GridLogPerformance;
+extern GridLogger GridLogDslash;
 extern GridLogger GridLogIterative  ;
 extern GridLogger GridLogIntegrator  ;
+extern GridLogger GridLogHMC;
+extern GridLogger GridLogMemory;
+extern GridLogger GridLogTracing;
 extern Colours    GridLogColours;
 
 std::string demangle(const char* name) ;

Grid/perfmon/PerfCount.cc

@@ -27,10 +27,13 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 /*  END LEGAL */
 
 #include <Grid/GridCore.h>
-#include <Grid/perfmon/PerfCount.h>
 
+#include <Grid/perfmon/Timer.h>
+#include <Grid/perfmon/PerfCount.h>
 NAMESPACE_BEGIN(Grid);
 
+GridTimePoint theProgramStart = GridClock::now();
+
 #define CacheControl(L,O,R) ((PERF_COUNT_HW_CACHE_##L)|(PERF_COUNT_HW_CACHE_OP_##O<<8)| (PERF_COUNT_HW_CACHE_RESULT_##R<<16))
 #define RawConfig(A,B) (A<<8|B)
 const PerformanceCounter::PerformanceCounterConfig PerformanceCounter::PerformanceCounterConfigs [] = {

Grid/perfmon/PerfCount.h

@@ -30,6 +30,12 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_PERFCOUNT_H
 #define GRID_PERFCOUNT_H
 
+
+#ifndef __SSC_START
+#define __SSC_START
+#define __SSC_STOP
+#endif
+
 #include <sys/time.h>
 #include <ctime>
 #include <chrono>
@@ -72,17 +78,9 @@ static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
 inline uint64_t cyclecount(void){ 
   return 0;
 }
-#define __SSC_MARK(mark) __asm__ __volatile__ ("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(mark):"%ebx")
-#define __SSC_STOP  __SSC_MARK(0x110)
-#define __SSC_START __SSC_MARK(0x111)
-
 
 #else
 
-#define __SSC_MARK(mark) 
-#define __SSC_STOP  
-#define __SSC_START 
-
 /*
  * cycle counters arch dependent
  */

Grid/perfmon/Timer.h

@@ -35,17 +35,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 NAMESPACE_BEGIN(Grid)
 
-// Dress the output; use std::chrono
+//typedef  std::chrono::system_clock          GridClock;
-// C++11 time facilities better?
+typedef  std::chrono::high_resolution_clock   GridClock;
-inline double usecond(void) {
-  struct timeval tv;
-#ifdef TIMERS_ON
-  gettimeofday(&tv,NULL);
-#endif
-  return 1.0*tv.tv_usec + 1.0e6*tv.tv_sec;
-}
-
-typedef  std::chrono::system_clock          GridClock;
 typedef  std::chrono::time_point<GridClock> GridTimePoint;
 
 typedef  std::chrono::seconds               GridSecs;
@@ -53,6 +44,15 @@ typedef  std::chrono::milliseconds          GridMillisecs;
 typedef  std::chrono::microseconds          GridUsecs;
 typedef  std::chrono::microseconds          GridTime;
 
+extern GridTimePoint theProgramStart;
+// Dress the output; use std::chrono
+// C++11 time facilities better?
+inline double usecond(void) {
+  auto usecs = std::chrono::duration_cast<GridUsecs>(GridClock::now()-theProgramStart); 
+  return 1.0*usecs.count();
+}
+
+
 inline std::ostream& operator<< (std::ostream & stream, const GridSecs & time)
 {
   stream << time.count()<<" s";

Grid/perfmon/Tracing.h

@@ -0,0 +1,70 @@
+#pragma once
+
+NAMESPACE_BEGIN(Grid);
+
+#ifdef GRID_TRACING_NVTX
+#include <nvToolsExt.h>
+class GridTracer {
+public:
+  GridTracer(const char* name) {
+    nvtxRangePushA(name);
+  }
+  ~GridTracer() {
+    nvtxRangePop();
+  }
+};
+inline void tracePush(const char *name) { nvtxRangePushA(name); }
+inline void tracePop(const char *name) { nvtxRangePop(); }
+inline int  traceStart(const char *name) {  }
+inline void traceStop(int ID) {  }
+#endif
+
+#ifdef GRID_TRACING_ROCTX
+#include <roctracer/roctx.h>
+class GridTracer {
+ public:
+  GridTracer(const char* name) {
+    roctxRangePushA(name);
+    std::cout << "roctxRangePush "<<name<<std::endl;
+  }
+  ~GridTracer() {
+    roctxRangePop();
+    std::cout << "roctxRangePop "<<std::endl;
+  }
+};
+inline void tracePush(const char *name) { roctxRangePushA(name); }
+inline void tracePop(const char *name) { roctxRangePop(); }
+inline int  traceStart(const char *name) { roctxRangeStart(name); }
+inline void traceStop(int ID) { roctxRangeStop(ID); }
+#endif
+
+#ifdef GRID_TRACING_TIMER
+class GridTracer {
+ public:
+  const char *name;
+  double elapsed;
+  GridTracer(const char* _name) {
+    name = _name;
+    elapsed=-usecond();
+  }
+  ~GridTracer() {
+    elapsed+=usecond();
+    std::cout << GridLogTracing << name << " took " <<elapsed<< " us" <<std::endl;
+  }
+};
+inline void tracePush(const char *name) {  }
+inline void tracePop(const char *name) {  }
+inline int  traceStart(const char *name) { return 0; }
+inline void traceStop(int ID) {  }
+#endif
+
+#ifdef GRID_TRACING_NONE
+#define GRID_TRACE(name) 
+inline void tracePush(const char *name) {  }
+inline void tracePop(const char *name) {  }
+inline int  traceStart(const char *name) { return 0;  }
+inline void traceStop(int ID) {  }
+#else
+#define GRID_TRACE(name) GridTracer uniq_name_using_macros##__COUNTER__(name);
+#endif
+NAMESPACE_END(Grid);

Grid/pugixml/pugixml.cc

@@ -16,7 +16,7 @@
 
 #ifdef __NVCC__
 #pragma push
-#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
+#ifdef __NVCC_DIAG_PRAGMA_SUPPORT__
 #pragma nv_diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
 #else
 #pragma diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"

Grid/qcd/QCD.h

@@ -456,9 +456,9 @@ template <class Fimpl>
 void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c)
 {
 #ifdef FAST_FERM_TO_PROP
-  autoView(p_v,p,AcceleratorWrite);
+  autoView(p_v,p,CpuWrite);
-  autoView(f_v,f,AcceleratorRead);
+  autoView(f_v,f,CpuRead);
-  accelerator_for(idx,p_v.oSites(),1,{
+  thread_for(idx,p_v.oSites(),{
       for(int ss = 0; ss < Ns; ++ss) {
       for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
 	p_v[idx]()(ss,s)(cc,c) = f_v[idx]()(ss)(cc); // Propagator sink index is LEFT, suitable for left mult by gauge link (e.g.)
@@ -484,9 +484,9 @@ template <class Fimpl>
 void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::PropagatorField &p, const int s, const int c)
 {
 #ifdef FAST_FERM_TO_PROP
-  autoView(p_v,p,AcceleratorWrite);
+  autoView(p_v,p,CpuRead);
-  autoView(f_v,f,AcceleratorRead);
+  autoView(f_v,f,CpuWrite);
-  accelerator_for(idx,p_v.oSites(),1,{
+  thread_for(idx,p_v.oSites(),{
       for(int ss = 0; ss < Ns; ++ss) {
       for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
 	f_v[idx]()(ss)(cc) = p_v[idx]()(ss,s)(cc,c); // LEFT index is copied across for s,c right index

Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h

@@ -459,11 +459,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 
    if( interior && exterior ) {
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSite); return;}
-#ifdef SYCL_HACK     
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL_TMP(HandDhopSiteSycl);    return; }
-#else
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSite);    return;}
-#endif     
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSite);    return;}
 #endif
@@ -474,6 +470,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteInt);    return;}
 #endif
    } else if( exterior ) {
+     acceleratorFenceComputeStream();
      if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteExt); return;}
      if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteExt);    return;}
 #ifndef GRID_CUDA
@@ -498,10 +495,9 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st,  DoubledGaugeField
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDag);     return;}
 #endif
-     acceleratorFenceComputeStream();
    } else if( interior ) {
-     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
+     if (Opt == WilsonKernelsStatic::OptGeneric    ) { KERNEL_CALLNB(GenericDhopSiteDagInt); return;}
-     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt);    return;}
+     if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALLNB(HandDhopSiteDagInt);    return;}
 #ifndef GRID_CUDA
      if (Opt == WilsonKernelsStatic::OptInlineAsm  ) {  ASM_CALL(AsmDhopSiteDagInt);     return;}
 #endif

Grid/stencil/Stencil.h

@@ -398,6 +398,8 @@ public:
   ////////////////////////////////////////////////////////////////////////
   void CommunicateBegin(std::vector<std::vector<CommsRequest_t> > &reqs)
   {
+    // Buffers are gathered AND synchronised
+    // Copies are MPI ISend OR asynch copy on copy stream
     reqs.resize(Packets.size());
     commtime-=usecond();
     for(int i=0;i<Packets.size();i++){
@@ -410,14 +412,18 @@ public:
       comms_bytes+=bytes;
       shm_bytes  +=2*Packets[i].bytes-bytes;
     }
-    _grid->StencilBarrier();// Synch shared memory on a single nodes
   }
 
   void CommunicateComplete(std::vector<std::vector<CommsRequest_t> > &reqs)
   {
+    // complete intranode
+    acceleratorCopySynchronise();
+    // complete MPI
     for(int i=0;i<Packets.size();i++){
       _grid->StencilSendToRecvFromComplete(reqs[i],i);
     }
+    // Everyone agrees we are all done
+    _grid->StencilBarrier(); 
     commtime+=usecond();
   }
   ////////////////////////////////////////////////////////////////////////
@@ -425,34 +431,10 @@ public:
   ////////////////////////////////////////////////////////////////////////
   void Communicate(void)
   {
-    if ( 0 ){
-      thread_region {
-	// must be called in parallel region
-	int mythread  = thread_num();
-	int maxthreads= thread_max();
-	int nthreads = CartesianCommunicator::nCommThreads;
-	assert(nthreads <= maxthreads);
-	if (nthreads == -1) nthreads = 1;
-	if (mythread < nthreads) {
-	  for (int i = mythread; i < Packets.size(); i += nthreads) {
-	    double start = usecond();
-	    uint64_t bytes= _grid->StencilSendToRecvFrom(Packets[i].send_buf,
-							 Packets[i].to_rank,
-							 Packets[i].recv_buf,
-							 Packets[i].from_rank,
-							 Packets[i].bytes,i);
-	    comm_bytes_thr[mythread] += bytes;
-	    shm_bytes_thr[mythread]  += Packets[i].bytes - bytes;
-	    comm_time_thr[mythread]  += usecond() - start;
-	  }
-	}
-      }
-    } else { // Concurrent and non-threaded asynch calls to MPI
     std::vector<std::vector<CommsRequest_t> > reqs;
     this->CommunicateBegin(reqs);
     this->CommunicateComplete(reqs);
   }
-  }
 
   template<class compressor> void HaloExchange(const Lattice<vobj> &source,compressor &compress)
   {
@@ -527,7 +509,6 @@ public:
     _grid->StencilBarrier();// Synch shared memory on a single nodes
     mpi3synctime_g+=usecond();
 
-    // conformable(source.Grid(),_grid);
     assert(source.Grid()==_grid);
     halogtime-=usecond();
 
@@ -586,13 +567,8 @@ public:
     CommsMerge(decompress,Mergers,Decompressions);
   }
   template<class decompressor>  void CommsMergeSHM(decompressor decompress) {
-    mpi3synctime-=usecond();
+    assert(MergersSHM.size()==0);
-    accelerator_barrier();
+    assert(DecompressionsSHM.size()==0);
-    _grid->StencilBarrier();// Synch shared memory on a single nodes
-    mpi3synctime+=usecond();
-    shmmergetime-=usecond();
-    CommsMerge(decompress,MergersSHM,DecompressionsSHM);
-    shmmergetime+=usecond();
   }
 
   template<class decompressor>
@@ -609,6 +585,7 @@ public:
 	  decompress.Exchange(mp,vp0,vp1,type,o);
       });
     }
+    if ( mm.size() )    acceleratorFenceComputeStream();
     mergetime+=usecond();
 
     decompresstime-=usecond();
@@ -619,7 +596,9 @@ public:
 	decompress.Decompress(kp,mp,o);
       });
     }
+    if ( dd.size() )    acceleratorFenceComputeStream();
     decompresstime+=usecond();
+    
   }
   ////////////////////////////////////////
   // Set up routines

Grid/threads/Accelerator.h

@@ -249,14 +249,16 @@ inline int  acceleratorIsCommunicable(void *ptr)
 //////////////////////////////////////////////
 #ifdef GRID_SYCL
 NAMESPACE_END(Grid);
+#if 0
 #include <CL/sycl.hpp>
 #include <CL/sycl/usm.hpp>
-
-#define GRID_SYCL_LEVEL_ZERO_IPC
-
-#ifdef GRID_SYCL_LEVEL_ZERO_IPC
 #include <level_zero/ze_api.h>
 #include <CL/sycl/backend/level_zero.hpp>
+#else
+#include <sycl/CL/sycl.hpp>
+#include <sycl/usm.hpp>
+#include <level_zero/ze_api.h>
+#include <sycl/ext/oneapi/backend/level_zero.hpp>
 #endif
 NAMESPACE_BEGIN(Grid);
 

systems/PVC/benchmarks/run-1tile.sh

@@ -4,7 +4,7 @@
 #SBATCH -p QZ1J-ICX-PVC
 ##SBATCH -p QZ1J-SPR-PVC-2C
 
-source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
+#source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
 
 export NT=8
 

systems/PVC/benchmarks/run-2tile-mpi.sh

@@ -4,7 +4,7 @@
 
 #SBATCH -p QZ1J-ICX-PVC
 
-source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
+#source /nfs/site/home/paboylex/ATS/GridNew/Grid/systems/PVC-nightly/setup.sh
 
 export NT=16
 
@@ -19,16 +19,14 @@ export SYCL_DEVICE_FILTER=gpu,level_zero
 export I_MPI_OFFLOAD_CELL=tile
 export EnableImplicitScaling=0
 export EnableWalkerPartition=0
-export SYCL_PI_LEVEL_ZERO_DEVICE_SCOPE_EVENTS=1
+#export SYCL_PI_LEVEL_ZERO_DEVICE_SCOPE_EVENTS=1
-export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+#export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
 export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0
 
-for i in 0 
+for i in 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
 do
-mpiexec -launcher ssh -n 2 -host localhost  ./wrap4gpu.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads $NT  --shm-mpi 1  --device-mem 32768
+mpiexec -launcher ssh -n 2 -host localhost  ./wrap.sh ./Benchmark_dwf_fp32 --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads $NT  --shm-mpi 0  --device-mem 32768 > 1.1.1.2.log$i
-mpiexec -launcher ssh -n 2 -host localhost  ./wrap4gpu.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT  --shm-mpi 1  --device-mem 32768
+mpiexec -launcher ssh -n 2 -host localhost  ./wrap.sh ./Benchmark_dwf_fp32 --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT  --shm-mpi 0  --device-mem 32768 > 2.1.1.1.log$i 
 done
-#mpiexec -launcher ssh -n 2 -host localhost  ./wrap4gpu.sh ./Benchmark_halo --mpi 1.1.1.2 --grid 32.32.32.64 --accelerator-threads $NT  --shm-mpi 1 > halo.2tile.1x2.log
-#mpiexec -launcher ssh -n 2 -host localhost  ./wrap4gpu.sh ./Benchmark_halo --mpi 2.1.1.1 --grid 64.32.32.32 --accelerator-threads $NT  --shm-mpi 1 > halo.2tile.2x1.log
 
 

systems/PVC/benchmarks/wrap.sh

@@ -5,10 +5,10 @@ export ZE_AFFINITY_MASK=0.$MPI_LOCALRANKID
 echo Ranke $MPI_LOCALRANKID ZE_AFFINITY_MASK is $ZE_AFFINITY_MASK
 
 
-if [ $MPI_LOCALRANKID = "0" ] 
+#if [ $MPI_LOCALRANKID = "0" ] 
-then
+#then
 #  ~psteinbr/build_pti/ze_tracer -h $@
-  onetrace --chrome-device-timeline $@
+#  onetrace --chrome-device-timeline $@
-else
+#else
   $@
-fi
+#fi

systems/mac-arm/config-command-mpi

@@ -1 +1 @@
-CXX=mpicxx-openmpi-mp CXXFLAGS=-I/opt/local/include/ LDFLAGS=-L/opt/local/lib/ ../../configure --enable-simd=GEN --enable-debug --enable-comms=mpi
+CXX=mpicxx-openmpi-mp CXXFLAGS=-I/opt/local/include/ LDFLAGS=-L/opt/local/lib/ ../../configure --enable-simd=GEN --enable-debug --enable-comms=mpi --enable-unified=no

tests/core/Test_fft_matt.cc

@@ -0,0 +1,270 @@
+    /*************************************************************************************
+    grid` physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_cshift.cc
+
+    Copyright (C) 2015
+
+Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+
+using namespace Grid;
+
+Gamma::Algebra Gmu [] = {
+  Gamma::Algebra::GammaX,
+  Gamma::Algebra::GammaY,
+  Gamma::Algebra::GammaZ,
+  Gamma::Algebra::GammaT,
+  Gamma::Algebra::Gamma5
+};
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  int threads = GridThread::GetThreads();
+  std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
+
+  Coordinate latt_size   = GridDefaultLatt();
+  Coordinate simd_layout = GridDefaultSimd(Nd,vComplexD::Nsimd());
+  Coordinate mpi_layout  = GridDefaultMpi();
+
+  int vol = 1;
+  for(int d=0;d<latt_size.size();d++){
+    vol = vol * latt_size[d];
+  }
+  GridCartesian         GRID(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian RBGRID(&GRID);
+
+  LatticeComplexD    coor(&GRID);
+  ComplexD ci(0.0,1.0);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridSerialRNG          sRNG;  sRNG.SeedFixedIntegers(seeds); // naughty seeding
+  GridParallelRNG          pRNG(&GRID);
+  pRNG.SeedFixedIntegers(seeds);
+
+  LatticeGaugeFieldD Umu(&GRID);
+  SU<Nc>::ColdConfiguration(pRNG,Umu); // Unit gauge
+
+  ////////////////////////////////////////////////////
+  // Wilson test
+  ////////////////////////////////////////////////////
+  {
+    LatticeFermionD    src(&GRID); gaussian(pRNG,src);
+    LatticeFermionD    src_p(&GRID);
+    LatticeFermionD    tmp(&GRID);
+    LatticeFermionD    ref(&GRID);
+    LatticeFermionD    result(&GRID);
+    
+    RealD mass=0.1;
+    WilsonFermionD Dw(Umu,GRID,RBGRID,mass);
+    
+    Dw.M(src,ref);
+    std::cout << "Norm src "<<norm2(src)<<std::endl;
+    std::cout << "Norm Dw x src "<<norm2(ref)<<std::endl;
+    {
+      FFT theFFT(&GRID);
+
+      ////////////////
+      // operator in Fourier space
+      ////////////////
+      tmp =ref;
+      theFFT.FFT_all_dim(result,tmp,FFT::forward);
+      std::cout<<"FFT[ Dw x src ]  "<< norm2(result)<<std::endl;    
+
+      tmp = src;
+      theFFT.FFT_all_dim(src_p,tmp,FFT::forward);
+      std::cout<<"FFT[ src      ]  "<< norm2(src_p)<<std::endl;
+      
+      /////////////////////////////////////////////////////////////////
+      // work out the predicted FT from Fourier
+      /////////////////////////////////////////////////////////////////
+      auto FGrid = &GRID;
+      LatticeFermionD    Kinetic(FGrid); Kinetic = Zero();
+      LatticeComplexD    kmu(FGrid); 
+      LatticeInteger     scoor(FGrid); 
+      LatticeComplexD    sk (FGrid); sk = Zero();
+      LatticeComplexD    sk2(FGrid); sk2= Zero();
+      LatticeComplexD    W(FGrid); W= Zero();
+      LatticeComplexD    one(FGrid); one =ComplexD(1.0,0.0);
+      ComplexD ci(0.0,1.0);
+    
+      for(int mu=0;mu<Nd;mu++) {
+	
+	RealD TwoPiL =  M_PI * 2.0/ latt_size[mu];
+
+	LatticeCoordinate(kmu,mu);
+
+	kmu = TwoPiL * kmu;
+      
+	sk2 = sk2 + 2.0*sin(kmu*0.5)*sin(kmu*0.5);
+	sk  = sk  +     sin(kmu)    *sin(kmu); 
+      
+	// -1/2 Dw ->  1/2 gmu (eip - emip) = i sinp gmu
+	Kinetic = Kinetic + sin(kmu)*ci*(Gamma(Gmu[mu])*src_p);
+	
+      }
+    
+      W = mass + sk2; 
+      Kinetic = Kinetic + W * src_p;
+    
+      std::cout<<"Momentum space src         "<< norm2(src_p)<<std::endl;
+      std::cout<<"Momentum space Dw x src    "<< norm2(Kinetic)<<std::endl;
+      std::cout<<"FT[Coordinate space Dw]    "<< norm2(result)<<std::endl;
+    
+      result = result - Kinetic;
+      std::cout<<"diff "<< norm2(result)<<std::endl;
+      
+    }
+
+    std::cout << " =======================================" <<std::endl;
+    std::cout << " Checking FourierFreePropagator x Dw = 1" <<std::endl;
+    std::cout << " =======================================" <<std::endl;
+    std::cout << "Dw src = " <<norm2(src)<<std::endl;
+    std::cout << "Dw tmp = " <<norm2(tmp)<<std::endl;
+    Dw.M(src,tmp);
+    Dw.FreePropagator(tmp,ref,mass);
+
+    std::cout << "Dw ref = " <<norm2(ref)<<std::endl;
+    
+    ref = ref - src;
+    
+    std::cout << "Dw ref-src = " <<norm2(ref)<<std::endl;
+  }
+
+
+  ////////////////////////////////////////////////////
+  // Wilson prop
+  ////////////////////////////////////////////////////
+  {
+    std::cout<<"****************************************"<<std::endl;
+    std::cout << "Wilson Mom space 4d propagator \n";
+    std::cout<<"****************************************"<<std::endl;
+
+    LatticeFermionD    src(&GRID); gaussian(pRNG,src);
+    LatticeFermionD    tmp(&GRID);
+    LatticeFermionD    ref(&GRID);
+    LatticeFermionD    diff(&GRID);
+
+    src=Zero();
+    Coordinate point(4,0); // 0,0,0,0
+    SpinColourVectorD ferm;
+    ferm=Zero();
+    ferm()(0)(0) = ComplexD(1.0);
+    pokeSite(ferm,src,point);
+
+    RealD mass=0.1;
+    WilsonFermionD Dw(Umu,GRID,RBGRID,mass);
+
+    // Momentum space prop
+    std::cout << " Solving by FFT and Feynman rules" <<std::endl;
+    Dw.FreePropagator(src,ref,mass) ;
+
+    Gamma G5(Gamma::Algebra::Gamma5);
+
+    LatticeFermionD    result(&GRID); 
+    const int sdir=0;
+    
+    ////////////////////////////////////////////////////////////////////////
+    // Conjugate gradient on normal equations system
+    ////////////////////////////////////////////////////////////////////////
+    std::cout << " Solving by Conjugate Gradient (CGNE)" <<std::endl;
+    Dw.Mdag(src,tmp);
+    src=tmp;
+    MdagMLinearOperator<WilsonFermionD,LatticeFermionD> HermOp(Dw);
+    ConjugateGradient<LatticeFermionD> CG(1.0e-10,10000);
+    CG(HermOp,src,result);
+    
+    ////////////////////////////////////////////////////////////////////////
+    std::cout << " Taking difference" <<std::endl;
+    std::cout << "Dw result "<<norm2(result)<<std::endl;
+    std::cout << "Dw ref     "<<norm2(ref)<<std::endl;
+    
+    diff = ref - result;
+    std::cout << "result - ref     "<<norm2(diff)<<std::endl;
+
+    DumpSliceNorm("Slice Norm Solution ",result,Nd-1);
+  }
+
+  ////////////////////////////////////////////////////
+  //Gauge invariance test
+  ////////////////////////////////////////////////////
+  {
+    std::cout<<"****************************************"<<std::endl;
+    std::cout << "Gauge invariance test \n";
+    std::cout<<"****************************************"<<std::endl;
+    LatticeGaugeField     U_GT(&GRID); // Gauge transformed field
+    LatticeColourMatrix   g(&GRID);    // local Gauge xform matrix
+    U_GT = Umu;
+    // Make a random xform to teh gauge field
+    SU<Nc>::RandomGaugeTransform(pRNG,U_GT,g); // Unit gauge
+
+    LatticeFermionD    src(&GRID);
+    LatticeFermionD    tmp(&GRID);
+    LatticeFermionD    ref(&GRID);
+    LatticeFermionD    diff(&GRID);
+
+    // could loop over colors
+    src=Zero();
+    Coordinate point(4,0); // 0,0,0,0
+    SpinColourVectorD ferm;
+    ferm=Zero();
+    ferm()(0)(0) = ComplexD(1.0);
+    pokeSite(ferm,src,point);
+
+    RealD mass=0.1;
+    WilsonFermionD Dw(U_GT,GRID,RBGRID,mass);
+
+    // Momentum space prop
+    std::cout << " Solving by FFT and Feynman rules" <<std::endl;
+    Dw.FreePropagator(src,ref,mass) ;
+
+    Gamma G5(Gamma::Algebra::Gamma5);
+
+    LatticeFermionD    result(&GRID); 
+    const int sdir=0;
+    
+    ////////////////////////////////////////////////////////////////////////
+    // Conjugate gradient on normal equations system
+    ////////////////////////////////////////////////////////////////////////
+    std::cout << " Solving by Conjugate Gradient (CGNE)" <<std::endl;
+    Dw.Mdag(src,tmp);
+    src=tmp;
+    MdagMLinearOperator<WilsonFermionD,LatticeFermionD> HermOp(Dw);
+    ConjugateGradient<LatticeFermionD> CG(1.0e-10,10000);
+    CG(HermOp,src,result);
+    
+    ////////////////////////////////////////////////////////////////////////
+    std::cout << " Taking difference" <<std::endl;
+    std::cout << "Dw result "<<norm2(result)<<std::endl;
+    std::cout << "Dw ref     "<<norm2(ref)<<std::endl;
+    
+    diff = ref - result;
+    std::cout << "result - ref     "<<norm2(diff)<<std::endl;
+
+    DumpSliceNorm("Slice Norm Solution ",result,Nd-1);
+  }
+  
+  
+  Grid_finalize();
+}

tests/core/Test_memory_manager.cc

@@ -0,0 +1,110 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_memory_manager.cc
+
+    Copyright (C) 2022
+
+Author: Peter Boyle <pboyle@bnl.gov>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include <Grid/Grid.h>
+
+using namespace std;
+using namespace Grid;
+
+void  MemoryTest(GridCartesian         * FGrid,int N);
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
+
+  int N=100;
+  for(int i=0;i<N;i++){
+    std::cout << "============================"<<std::endl;
+    std::cout << "Epoch "<<i<<"/"<<N<<std::endl;
+    std::cout << "============================"<<std::endl;
+    MemoryTest(UGrid,256);
+    MemoryManager::Print();
+    AUDIT();
+  }
+  Grid_finalize();
+}
+
+void  MemoryTest(GridCartesian         * FGrid, int N)
+{
+  LatticeComplexD zero(FGrid); zero=Zero();
+  std::vector<LatticeComplexD> A(N,zero);//FGrid);
+
+  std::vector<ComplexD> B(N,ComplexD(0.0)); // Update sequentially on host
+
+  for(int v=0;v<N;v++) A[v] = Zero();
+
+  uint64_t counter = 0;
+  for(int epoch = 0;epoch<10000;epoch++){
+
+    int v  = random() %N; // Which vec
+    int w  = random() %2; // Write or read
+    int e  = random() %3; // expression or for loop
+    int dev= random() %2; // On device?
+    //    int e=1;
+    ComplexD zc = counter++;
+    
+    if ( w ) {
+      B[v] = B[v] + zc;
+      if ( e == 0 ) {
+	A[v] = A[v] + zc - A[v] + A[v];
+      } else {
+	if ( dev ) { 
+	  autoView(A_v,A[v],AcceleratorWrite);
+	  accelerator_for(ss,FGrid->oSites(),1,{
+	    A_v[ss] = A_v[ss] + zc;
+	    });
+	} else {
+	  autoView(A_v,A[v],CpuWrite);
+	  thread_for(ss,FGrid->oSites(),{
+	      A_v[ss] = A_v[ss] + zc;
+	    });
+	}
+      }
+    } else {
+      if ( e == 0 ) {
+	A[v] = A[v] + A[v] - A[v];
+      } else { 
+	if ( dev ) { 
+	  autoView(A_v,A[v],AcceleratorRead);
+	  accelerator_for(ss,FGrid->oSites(),1,{
+	      assert(B[v]==A_v[ss]()()().getlane(0));
+	    });
+	  //	std::cout << "["<<v<<"] checked on GPU"<<B[v]<<std::endl;
+	} else {
+	  autoView(A_v,A[v],CpuRead);
+	  thread_for(ss,FGrid->oSites(),{
+	      assert(B[v]==A_v[ss]()()().getlane(0));
+	    });
+	  //	std::cout << "["<<v<<"] checked on CPU"<<B[v]<<std::endl;
+	}
+      }    
+    }
+  }
+
+}